feat: metal backend

[dcvz]

No description provided.

[Copilot]

Pull request overview

Adds a native Metal backend to CubeCL and wires it into the workspace’s build/test/doc tooling and codegen so Metal can be selected as a runtime and supported in CI workflows.

Changes:

• Introduces new cubecl-metal crate implementing a Metal runtime/server/stream/memory/storage stack.
• Exposes Metal runtime via cubecl feature flags (metal) and TestRuntime cfg plumbing.
• Updates xtask + CI workflow to support doc --ci and to exclude Metal/CUDA/HIP crates in CI.

Reviewed changes

Copilot reviewed 32 out of 32 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
xtask/src/main.rs	Adds doc subcommand wiring to the xtask CLI.
xtask/src/commands/mod.rs	Registers new doc module.
xtask/src/commands/doc.rs	Adds --ci handling for docs, excluding unsupported crates.
xtask/src/commands/validate.rs	Routes doc validation through the new CubeCL doc command args.
xtask/src/commands/build.rs	Excludes cubecl-metal (and CUDA/HIP) in CI builds.
xtask/src/commands/test.rs	Excludes cubecl-metal (and CUDA/HIP) in CI tests.
xtask/src/commands/check.rs	Adds CI-specific workspace clippy that excludes platform-specific crates.
.github/workflows/ci.yml	Switches CI invocations to use --ci for check and doc.
crates/cubecl/src/lib.rs	Re-exports cubecl_metal behind feature = "metal" and adds test_runtime_metal.
crates/cubecl/Cargo.toml	Re-points metal feature to cubecl-metal and adds the dependency.
crates/cubecl/build.rs	Adds test_runtime_metal check-cfg and feature wiring.
crates/cubecl-metal/*	New Metal runtime implementation (device selection, compilation, server, stream/event sync, storage).
crates/cubecl-cpp/src/shared/variable.rs	Updates pointer/atomic formatting to include Metal address spaces.
crates/cubecl-cpp/src/shared/unary.rs	Adjusts unary function formatting for BF16 under Metal math constraints.
crates/cubecl-cpp/src/shared/instruction.rs	Casts float literals to support BF16 paths.
crates/cubecl-cpp/src/shared/base.rs	Ensures Metal extensions get registered for hypot/rhypot.
crates/cubecl-cpp/src/metal/extension.rs	Adds hypot/rhypot extensions and improves BF16-safe casts.
crates/cubecl-cpp/src/metal/dialect.rs	Emits hypot/rhypot calls and fixes atomic compare-exchange emission for MSL semantics.
crates/cubecl-cpp/src/metal/arch.rs	Adjusts Metal “warp” size to 32.
crates/cubecl-cpp/src/metal/address_space.rs	Ensures atomic bindings/pointers use device address space in MSL.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

enumerate_devices ignores the type_id filter and always returns DeviceId { type_id: 0, ... }. With the current MetalDevice::from_id mapping, this means every enumerated device resolves to DefaultDevice, and callers can’t enumerate/select discrete vs integrated devices correctly. Consider implementing type_id filtering similar to cubecl-wgpu (e.g., return only DefaultDevice for type 0, discrete list for type 1, integrated list for type 2, etc.), and ensure returned DeviceId.type_id matches the MetalDevice variants you support.

Suggested change

	_type_id: u16,
	_info: &<Self::Server as cubecl_core::server::ComputeServer>::Info,
	) -> Vec<DeviceId> {
	let devices = crate::device::all_devices();
	(0..devices.len())
	.map(|i| DeviceId {
	type_id: 0,
	index_id: i as u32,
	})
	.collect()
	type_id: u16,
	_info: &<Self::Server as cubecl_core::server::ComputeServer>::Info,
	) -> Vec<DeviceId> {
	let devices = crate::device::all_devices();
	match type_id {
	0 => vec![DeviceId {
	type_id: 0,
	index_id: 0,
	}],
	1 => devices
	.iter()
	.enumerate()
	.filter_map(|(i, device)| {
	(!device.is_low_power()).then_some(DeviceId {
	type_id: 1,
	index_id: i as u32,
	})
	})
	.collect(),
	2 => devices
	.iter()
	.enumerate()
	.filter_map(|(i, device)| {
	device.is_low_power().then_some(DeviceId {
	type_id: 2,
	index_id: i as u32,
	})
	})
	.collect(),
	_ => Vec::new(),
	}

[Copilot]

MultiStream::resolve is called only with bindings.buffers.iter(). When count is CubeCount::Dynamic, the indirect-dispatch Binding used for dispatchThreadgroupsWithIndirectBuffer... is not included in the resolve handles, so cross-stream synchronization for that buffer can be skipped. Consider including the dynamic dispatch Binding in the iterator passed to resolve (e.g., chain the buffers iterator with std::iter::once(binding) when CubeCount::Dynamic).

Suggested change

	let dispatch_info = match count {
	CubeCount::Static(x, y, z) => DispatchInfo::Static(x, y, z),
	CubeCount::Dynamic(binding) => DispatchInfo::Dynamic(binding),
	};
	let mut resolved = match self
	.streams
	.resolve(stream_id, bindings.buffers.iter(), false)
	{
	Ok(r) => r,
	Err(_) => return,
	};
	let mut resolved = match &count {
	CubeCount::Dynamic(binding) => self.streams.resolve(
	stream_id,
	bindings
	.buffers
	.iter()
	.chain(std::iter::once(binding)),
	false,
	),
	CubeCount::Static(_, _, _) => {
	self.streams.resolve(stream_id, bindings.buffers.iter(), false)
	}
	} {
	Ok(r) => r,
	Err(_) => return,
	};
	let dispatch_info = match count {
	CubeCount::Static(x, y, z) => DispatchInfo::Static(x, y, z),
	CubeCount::Dynamic(binding) => DispatchInfo::Dynamic(binding),
	};

[Copilot]

handle_cursor currently returns handle.size, which is constant for the lifetime of an allocation and does not reflect when the binding was last written on a stream. This breaks MultiStream’s shared-binding synchronization logic (it relies on a monotonic cursor, like other backends’ memory_management.get_cursor(binding.memory)), and can lead to missing waits between streams. Consider returning the cursor from stream.memory_management.get_cursor(handle.memory.clone()) (and falling back safely if that errors).

Suggested change

	fn handle_cursor(_stream: &Self::Stream, handle: &Binding) -> u64 {
	// Metal uses shared memory so cursor tracking is minimal
	handle.size
	fn handle_cursor(stream: &Self::Stream, handle: &Binding) -> u64 {
	stream
	.memory_management
	.get_cursor(handle.memory.clone())
	.unwrap_or(handle.size)

[Copilot]

cubecl-metal is added as an optional dependency without default-features = false, unlike the other runtime dependencies in this crate (cubecl-cuda, cubecl-hip, cubecl-wgpu, etc.). This can unintentionally pull in cubecl-metal default features when consumers enable cubecl/metal, which may be inconsistent with the crate’s feature strategy. Consider adding default-features = false for cubecl-metal for parity with the other runtime deps.

Suggested change

	cubecl-metal = { path = "../cubecl-metal", version = "=0.10.0-pre.3", optional = true }
	cubecl-metal = { path = "../cubecl-metal", version = "=0.10.0-pre.3", default-features = false, optional = true }